Quantization

Compressing a model by lowering the precision of its internal numbers (from 16 bits to 8, 4, or even fewer) so it fits on a more modest machine and runs faster. It is why you can run an open-weight model on your laptop instead of a multi-GPU server. The trade-off: at a given compression, the model stays very good on everyday tasks but loses a bit of sharpness on hard reasoning and precise details.

Strengths

Limitations

Best for

Official site

View on Coeurdar